DS-Project-Final.knit

Introduction

Your weather app doesn’t always match what you see when you look out the window. Despite considerable advances in weather reporting, the National Weather Service (NWS) — the primary data source for all weather service providers in the United States — struggles to maintain comparable accuracy and precision during periods of high volatility. Moreover, satellites, radars, and other forecasting tools synthesize complex weather patterns across large geographic areas and are not well suited for observing sudden changes in surface-level precipitation. In fact, radar – the NWS’s primary tool for measuring precipitation observed on the ground – does not measure surface-level rainfall at all. Meteorologists infer how much it’s raining on the ground by emitting electromagnetic energy and analyzing the “echo” that precipitation particles reflect back.

United States National Radar

radar

NATIONAL WEATHER SERVICE | Radar

While this approach can provide valuable information about overall precipitation patterns, detecting sudden shifts in local weather events requires eyes on the ground. As such, the NWS must rely on citizen volunteers to understand timely changes in severe weather condition and daily precipitation observed at the surface.

Using radar to estimate observed weather conditions isn’t exactly a Watergate-level scandal (especially considering I learned this information from NWS website). However, whether the NWS can provide timely, local weather warnings during periods of increased weather volatility hinges on its ability to accurately assess conditions on the ground. This mandate has incredible implications here in Washington, D.C., where urban topography and impervious surfaces, along with rising temperatures induced by climate change, have increased rainfall volatility and interior flooding (Cone, 2012; DC Silver Jackets, 2017; Zahura & Goodall, 2022).

Using my own personal weather station (PWS) in Dupont Circle (Figure 1), I compare NWS hourly rainfall estimates to observed rainfall at home to assess whether NWS estimates for the greater D.C. area representative of precipitation observed at the neighborhood level.

Figure 1 - Ambient Weather WS-7078 Smart Weather Station

raingauge

Sensory Array
1 Antenna
2 Rain collector
3 UV / light sensor
4 Mounting pole
5 Mounting Base
6 Balance indicator
7 Wind cups
8 Radiation shield
9 Wind vane

Data and Methods

Weather Underground reports live and historical weather data from a worldwide public network of personal weather stations. To keep track of my station’s reported weather conditions, I registered it on wundergound.com on November 13 and began tracking outdoor weather conditions the following day.

Official data for the Washington, D.C. Metropolitan area are estimated via radar at the Washington/Baltimore regional weather station located at Reagan Nation Airport. Data are reported hourly and are not available beyond three days through any public source. That meant I had to grab data from their website at least once every three days to avoid any critical gaps in information. While I managed to accomplish this feat, the same cannot be said for my weather station at home.

Home rainfall data for November 27 are missing due to human error (i.e., my roommate mistakenly moving the station just under our awning per my admittedly ambiguous instructions). To remedy this shortcoming and mitigate other potential measurement errors of which I am not aware, I cross reference my home data with data from two other nearby personal weather stations in the Weather Underground network. The stations are located in Adams Morgan and just east of Dupont Circle.

Limitations

Unsurprisingly, this glorified science fair experiment came with some challenges and limitations. As alluded to before, radars and rain gauges measure rainfall through methods that are not apples to apples. While my home station and the Adams Morgan station have a rain gauge accuracy of ±7% and ±10% respectively, accuracy for the station east of Dupont and the radar at Reagan Nation Airport are unknown. Moreover, with limited information and quality control checks on on other personal weather stations set ups, its plausible that they are also vulnerable to a host of other third party factors compromising the accuracy of their reports.A summary of data and design limitations are thus presented in Table 1.

Table 1 – Study Limitations

Weather Station	NWS	Home	PWS1	PWS2
Station Location	Regean Airport	N of Dupont	E of Dupont	Adams Morgan
Station Type	Unknown	WS-7078	Unknown	WS-1400-IP
Method of Rain Detection	Radar	Rain Gauge	Rain Gauge	Rain Gauge
Position of Detected Rainfall	Closer to Clouds	Surface Level	Surface Level	Surface Level
Measurement Accuracy	Unknown	± 7%	Unknown	± 10%
Risk of Roommate/Squirrel Interference	Low	High	Medium	Medium

Procedure

I copied data from tables on the NWS and Weather Underground websites from November 14 - December 10 and placed them into excel for preliminary data cleaning. I then uploaded data into R where I filtered for hours in which at least one station recorded 0.01 inches of rain or more. After gathering descriptive statistics for hours in which it rained (Table 2), I compare recorded precipitation amounts between each station using a series of OLS linear regressions. I also ran regressions to assess whether variation in recorded temperature, pressure, and humidity is comperable to variation in measured precipitation.

Hypotheses

Although variation in measured rainfall between each station may be minimal, there is a statistically significant difference in rainfall reported by the NWS station and each of the other local stations.
Differences in measured rainfall between the three neighboring stations are smaller than between these stations and the NWS station.
Variation between my home station and the NWS station is greater for reported rainfall than for any other metric.

Results

There have been approximately seven rainy days since data collection began on November 14. Across those 7 days, there are 45 hours for which at least 0.01 inches of rain is recorded (that is 39 hours of recorded rainfall at home given data for November 27 are missing. Although this sample size is large enough for a statistical analysis, it is still quite small. Daily accumulated precipitation and the average among all stations are displayed in Table 2.

Table 2 – Daily Accumulated Precipitation (inches)

Date	NWS (DCA)	Home (Dupont)	East Dupont	Adams Morgan	Avg. Accum.
Nov 30	0.32	0.37	0.40	0.35	0.360
Nov 27	0.24	NA	0.23	0.20	0.222
Nov 25	0.15	0.17	0.17	0.19	0.170
Nov 15	1.22	2.56	0.85	1.39	1.502
Dec 7	0.05	0.01	0.02	0.02	0.025
Dec 6	0.17	0.15	0.21	0.19	0.179
Dec 3	0.31	0.31	0.37	0.32	0.326

Graphing measured precipitation in Figure 2 illustrates how, aside from data collected on November 15, measured rainfall between local stations hovers around the same amount. Granted, the visualization suggests that whether I keep or disregard outliers, particularly the two heighest values measured by the home station, may notably change the estimated difference in hourly rainfall between stations.

Figure 2 –Observed Hourly Precipitation (inches)

A quick boxplot reveals that there are five outliers ranging from approx. 0.22 - 0.8 that I would be statistically justified in removing. However, due to the small sample size, I elect to remove only the two data points above 0.6 inches.

Figure 3 – Box Plot Distribution of Hourly Rainfall

(Home Weather Station)

Like Figure 2, the line graph in Figure 4 represents stations’ measured hourly rainfall, excluding outlying values of 0.69 or above.

Figure 4 – Observed Hourly Precipitation (inches)

Inferential Results

According to the below t-test results, there is not a statistically significant difference between rainfall measured by the National Weather Station in Arlington, and rainfall measured at the three local personal weather stations. Thus, the first hypothesis is incorrect.

Table 3 – T-Test Results: NWS Station vs. Personal Weather Stations

NWS Average	Home Average	p-value
0.057	0.092	0.253

NWS Average	EDup Average	p-value
0.055	0.05	0.707

NWS Average	AdMo Average	p-value
0.055	0.059	0.769

After conducting a series of simple OLS regressions, it seems that the NWS station, my Home station, and the station East of Dupont Circle are all most correlated to the station in Adams Morgan. However, my Home station and the E. Dupont station are both more correlated to the NWS station than they are to each other. As such, the second hypothesis is also incorrect.

Table 4 – OLS Results: Comparing Rainfall Measured by all Stations

	NWS	NWS	NWS	Home	Home	Adams Morgan
Home	0.06 ***
	(0.01)
E. Dupont		0.06 ***		0.14 ***		0.07 ***
		(0.00)		(0.02)		(0.01)
Adams Morgan			0.06 ***		0.16 ***
			(0.00)		(0.01)
N	39	45	45	39	39	45
R2	0.76	0.79	0.93	0.66	0.85	0.82
All continuous predictors are mean-centered and scaled by 1 standard deviation. * p < 0.001; p < 0.01; * p < 0.05.

Finally, when comparing all aspects of the weather measured by both statiions, it appears that the temperature, dew point, pressure, and humidity measured by both stations is highly comparable, with data measured by one station accounting for 91% - 100% of the variation in data collected by the other. Given that rainfall measurd by the NWS station accounts for only 76% of the rainfall measured by the Home station, the final hypothesis is correct.

Table 5 – OLS Results: Comparing All Weather Measurements between the NWS and Home Station

	NWS Rain	NWS Temp	NWS Dew	NWS Pressure	NWS Humidiity
Home Rain	0.06 ***
	(0.01)
Home Temp		5.31 ***
		(0.18)
Home Dew			4.88 ***
			(0.23)
Home Pressure				0.14 ***
				(0.00)
Home Humidity					7.71 ***
					(0.41)
N	39	39	39	39	39
R2	0.76	0.96	0.92	1.00	0.91
All continuous predictors are mean-centered and scaled by 1 standard deviation. * p < 0.001; p < 0.01; * p < 0.05.

Conclusion

Results suggest that measured rainfall is not distinctly different between the NWS station at Reagan National Airport and the other personal stations in the Dupont Circle and Adams Morgan area. Granted, coming to that conclusion using these results alone may be jumping the gun, as there are many limitations to this study. Not only are the weather stations’ sample sizes quite small and measurement errors unknown, but I conduct this study in the Fall when temperatures are cooler and rainfall is generally less volatile.

I recommend that future studies use year round data, or at least data collected during warmer periods, as that could reveal a different picture about how well NWS weather station reports reflect weather reported on the surface. Still, it’s promising that these initial results suggest that NWS radar predictions may not too far off.